OpenAI published an addendum to the GPT-5 system card that explains an important change in how the model responds to situations of emotional distress. The October 3 update improves ChatGPT's default behavior to better recognize signs of suffering and guide people toward real‑world support.
What OpenAI announced
The note clarifies that, since the release of GPT-5, the team worked on better benchmarks and on strengthening safety in conversations related to mental health and emotional distress. On October 3 they deployed an update meant to help ChatGPT more reliably recognize signs of distress, respond more carefully, and point users to human and professional resources.
They collaborated with more than 170 mental health experts to refine the responses. And the results? According to OpenAI, responses that did not meet the desired behavior dropped between 65 and 80 percent.
How they evaluated the change
The addendum publishes baseline evaluations that compare the August 15 version, known as GPT-5 Instant, with the updated version released on October 3. The evaluations measure how well the model recognizes risk signals, offers empathy, and suggests concrete steps toward human help.
Reducing responses that fail by 65–80% sounds strong, but what does fail mean? In this context it includes not recognizing urgency, not offering guidance toward real support, or giving responses that minimize the problem. Working with experts helped define those criteria and create test examples.
What changes for users and developers
For you as a user: it’s more likely that, when there are signs of distress, ChatGPT will respond with empathy and offer concrete options — like suggesting professional help, pointing to local crisis lines, or proposing immediate safety steps. It’s not a substitute for a professional, but it can be a more reliable point of support in hard moments.
For developers and teams integrating GPT-5: the addendum and evaluation details serve as a reference to understand the model’s default limits and capabilities. If you build apps that handle sensitive topics, it remains crucial to keep human oversight, verify local resources, and test with real users.
Limitations and open questions
Not everything is solved. Models can still fail with cultural nuances, indirect expressions of distress, or adversarial inputs designed to confuse them. The 65–80% reduction does not mean zero errors. Effectiveness also varies by language, context, and interface design.
OpenAI says it will continue working on benchmarks and external evaluation. What should we expect? More transparency about evaluation criteria, independent audits, and iterative improvements that include cultural and linguistic diversity.
A practical example (simple)
Imagine someone writes that they no longer want to continue. With the update, the model will tend to:
- recognize urgency and express empathy;
- ask about immediate safety in a careful way;
- offer concrete resources, like helplines and a recommendation to contact professionals or emergency services when appropriate;
- avoid responses that minimize the situation.
This doesn’t replace a professional, but it reduces the chance of harmful or indifferent replies.
Final reflection
Seeing a major platform invest in collaboration with experts and publish evaluations is a positive sign. AI remains a tool: it can improve detection and guidance in critical moments, but it needs human accompaniment, responsible design, and constant iteration.
Original source
https://openai.com/index/gpt-5-system-card-sensitive-conversations